Anecdata seems quite valid for LLM comparison when trying to evaluate 'usefullness' for users. The lmsys chat leaderboard is literally just mass anecdata.